app.py

import gradio as gr
import torch
import os
from transformers import SpeechT5Processor, SpeechT5ForTextToSpeech, SpeechT5HifiGan
from datasets import load_dataset, Audio
import numpy as np
from speechbrain.inference import EncoderClassifier

# Load models and processor
processor = SpeechT5Processor.from_pretrained("microsoft/speecht5_tts")
model = SpeechT5ForTextToSpeech.from_pretrained("Tejasva-Maurya/Hindi_SpeechT5_finetuned")
vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan")

# Load speaker encoder
device = "cuda" if torch.cuda.is_available() else "cpu"
speaker_model = EncoderClassifier.from_hparams(
    source="speechbrain/spkrec-xvect-voxceleb",
    run_opts={"device": device},
    savedir=os.path.join("/tmp", "speechbrain/spkrec-xvect-voxceleb")
)

def create_speaker_embedding(waveform):
      with torch.no_grad():
          speaker_embeddings = speaker_model.encode_batch(torch.tensor(waveform))
          speaker_embeddings = torch.nn.functional.normalize(speaker_embeddings, dim=2)
          speaker_embeddings = speaker_embeddings.squeeze().cpu().numpy()
      return speaker_embeddings
def prepare_dataset(example):
    audio = example["audio"]
    example["speaker_embeddings"] = create_speaker_embedding(audio["array"])
    return example

# Load a sample from the dataset for speaker embedding
try:
  dataset = load_dataset("mozilla-foundation/common_voice_17_0", "hi", split="validated", trust_remote_code=True)
  dataset = dataset.cast_column("audio", Audio(sampling_rate=16000))
  spk_model_name = "speechbrain/spkrec-xvect-voxceleb"
  device = "cuda" if torch.cuda.is_available() else "cpu"
  speaker_model = EncoderClassifier.from_hparams(
      source=spk_model_name,
      run_opts={"device": device},
      savedir=os.path.join("/tmp", spk_model_name),
  )
  # Calculate the number of rows for a part of the dataset
  part = len(dataset) //800
  
  # Select the part of the dataset
  dataset = dataset.select(range(part))
  
  # Prepare the dataset
  dataset = dataset.map(prepare_dataset, remove_columns=dataset.column_names)
  example = dataset[5]
  speaker_embeddings = torch.tensor(example["speaker_embeddings"]).unsqueeze(0)
  
except Exception as e:
    print(f"Error loading dataset: {e}")
    # Use a random speaker embedding as fallback
    speaker_embedding = torch.randn(1, 512)

def text_to_speech(text):
    replacements = [
       # Vowels and vowel matras
       ("अ", "a"),
       ("आ", "aa"),
       ("इ", "i"),
       ("ई", "ee"),
       ("उ", "u"),
       ("ऊ", "oo"),
       ("ऋ", "ri"),
       ("ए", "e"),
       ("ऐ", "ai"),
       ("ऑ", "o"),  # More accurate than 'au' for ऑ
       ("ओ", "o"),
       ("औ", "au"),
       # Consonants
       ("क", "k"),
       ("ख", "kh"),
       ("ग", "g"),
       ("घ", "gh"),
       ("ङ", "ng"),  # nasal sound
       ("च", "ch"),
       ("छ", "chh"),
       ("ज", "j"),
       ("झ", "jh"),
       ("ञ", "ny"),  # 'ny' closer to the actual sound
       ("ट", "t"),
       ("ठ", "th"),
       ("ड", "d"),
       ("ढ", "dh"),
       ("ण", "n"),  # Slight improvement for easier pronunciation
       ("त", "t"),
       ("थ", "th"),
       ("द", "d"),
       ("ध", "dh"),
       ("न", "n"),
       ("प", "p"),
       ("फ", "ph"),
       ("ब", "b"),
       ("भ", "bh"),
       ("म", "m"),
       ("य", "y"),
       ("र", "r"),
       ("ल", "l"),
       ("व", "v"),  # 'v' is closer to the Hindi 'व'
       ("श", "sh"),
       ("ष", "sh"),  # Same sound in modern pronunciation
       ("स", "s"),
       ("ह", "h"),
       # Consonant clusters and special consonants
       ("क्ष", "ksh"),
       ("त्र", "tr"),
       ("ज्ञ", "gya"),
       ("श्र", "shra"),
       # Special characters
       ("़", ""),    # Ignore nukta; can vary with regional pronunciation
       ("्", ""),    # Halant - schwa dropping (handled contextually)
       ("ऽ", ""),    # Avagraha - no direct pronunciation, often ignored
       ("ं", "n"),   # Anusvara - nasalization
       ("ः", "h"),   # Visarga - adds an 'h' sound
       ("ँ", "n"),   # Chandrabindu - nasalization
       # Vowel matras (diacritic marks)
       ("ा", "a"),
       ("ि", "i"),
       ("ी", "ee"),
       ("ु", "u"),
       ("ू", "oo"),
       ("े", "e"),
       ("ै", "ai"),
       ("ो", "o"),
       ("ौ", "au"),
       ("ृ", "ri"),  # Vowel-matra equivalent of ऋ
       # Nasalization and other marks
       ("ॅ", "e"),   # Short 'e' sound (very rare)
       ("ॉ", "o"),   # Short 'o' sound (very rare)
       # Loanwords and aspirated consonants
       ("क़", "q"),
       ("ख़", "kh"),
       ("ग़", "gh"),
       ("ज़", "z"),
       ("ड़", "r"),
       ("ढ़", "rh"),
       ("फ़", "f"),
       # Punctuation
       ("।", "."),   # Hindi sentence-ending marker -> period
    ]

    # Remove extra whitespace
    text = ' '.join(text.split())
    for src, dst in replacements:
        text = text.replace(src, dst)
    
    inputs = processor(text=text, return_tensors="pt")
    speech = model.generate_speech(inputs["input_ids"], speaker_embedding, vocoder=vocoder)
    return (16000, speech.numpy())

iface = gr.Interface(
    fn=text_to_speech,
    inputs="text",
    outputs="audio",
    title="SpeechT5 finetuned Hindi Text-to-Speech",
    description="Enter Hindi text to convert it into an Audio"
)

iface.launch(share=True)